How AI helped improve beer without adding more alcohol

Eleven of thirteen tasters gave a modified non-alcoholic beer a higher overall appreciation score than its unmodified counterpart. No ethanol had been added. That is the result of AI News’ count of the paired scores in the researchers’ released data, and a useful place to begin understanding what AI can contribute to flavour.[2]

The experiment belongs to a Belgian study published in Nature Communications in March 2024. Researchers analysed 250 commercial beers, measured more than 200 chemical properties and assembled sensory assessments from a 16-person trained panel alongside about 180,000 online reviews. Their question was whether chemical composition could predict how a beer would be perceived—and suggest an improvement.[1]

The work took five years, according to VIB’s account. It began with a smaller collection of beers and expanded as the team recognised how much variation it needed to capture. Building that reference material required repeated chemical measurements and a trained vocabulary for describing flavour. The computer inherited a substantial piece of experimental craftsmanship.[3]

What the model was learning

A chemical profile is a set of measurements, not a recipe. It describes the resulting drink. That makes the task closer to learning which combinations tend to accompany a particular sensation than to asking a chatbot to invent an appealing list of ingredients.

The team trained on 175 beers and reserved 75 for testing. Gradient boosting, which combines many small decision trees, performed best overall. For overall appreciation, its test-set R² was 0.67 against online review scores but only 0.09 against the trained panel’s scores.[1]

R² compares a model’s prediction errors with those of a simple prediction based on the average. The two results therefore reveal something more useful than a single claim of “accuracy”: success depends on whose judgement the model is trying to predict. Learning a pattern in a large review collection does not amount to knowing what every drinker will enjoy.

The researchers then tested whether the patterns could help alter a drink. They added combinations of flavour-related compounds to existing beers and compared the results with controls in blind tastings. This moved the work from predicting an existing score to changing the liquid that would receive the score.[1]

The experiment without extra ethanol

The released data contain two particularly revealing comparisons: a blond beer modified without adding ethanol, and a non-alcoholic beer modified the same way. Their additions included glycerol and ethyl acetate. In each, 11 of 13 tasters scored the modified sample higher for overall appreciation and two scored it lower. There were no tied overall scores in either comparison. These are small tasting groups, not an estimate of how all beer drinkers would respond.[2]

The important idea is that a product developer can search for a preferred balance while holding one ingredient constant. If alcohol content is a design constraint, the optimisation problem becomes more interesting: which other changes can make the drink satisfying within that constraint? The experiment supplies an example of that approach working in a glass.

It also has a clear boundary. The authors found that their models did not reliably capture the point at which increasing a flavour compound becomes unpleasant, because commercial beers rarely supplied such extreme examples. A high prediction is therefore no instruction to keep adding more.[1]

Turning a promising chemical balance into a repeatable brewing process is a further task. The Verstrepen Lab describes work with industrial partners on yeast strains, fermentation and aroma formation, supported by a pilot brewery. Those are routes through which a brewer could try to produce a desired profile. The tasting experiment itself does not establish a finished commercial recipe.[4]

This is a productive role for AI in food: helping researchers make a more informed choice about which physical experiment deserves their time. The desired result remains wonderfully specific. A better non-alcoholic beer begins with a more precise idea of what, besides alcohol, people were enjoying.

Report an error or suggest a correction